arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2003.04710eess.AScs.SD
收藏

基于迁移学习的哈萨克语自动语音识别的开发

Development of Automatic Speech Recognition for Kazakh Language using Transfer Learning

Suleyman Demirel University(苏莱曼德米雷尔大学)

Amirgaliyev E. N., Kuanyshbay D. N., Baimuratov O

摘要:

哈萨克语自动语音识别系统的开发是一个很有挑战性的课题,由于数据的缺乏,现有的哈萨克语语音及其对应的音译词的数据访问量很大,不足以获得有价值的结果,因此哈萨克语的语音识别还没有得到很好的研究,只有很少的工作是用传统的方法,如隐马尔可夫模型,高斯混合模型,但他们的结果很差,缺乏足够的数据。在我们的工作中,我们提出了一种新的方法,采用预先训练的俄语模型,并将其知识作为我们的神经网络结构的起点,这意味着我们正在转移前的重量,我们选择俄语模型的主要原因是哈萨克语和俄语的发音非常相似,因为它们共享78%的字母,并且有相当大的俄语语音数据集。我们在苏莱曼·德米雷尔大学的基础上收集了一个哈萨克语带transmittance的语音数据集,50名母语者,每个人大约有400句话。我们在实验中考虑了四种不同的情况。首先,我们在没有使用预训练的俄罗斯模型的情况下训练了我们的神经网络,该模型具有2个LSTM层和2个BiLSTM。其次,我们使用预训练的模型训练了相同的2个LSTM层和2个BiLSTM层。因此,通过使用外部俄语语音识别模型,我们将模型的训练成本和标签错误率分别提高了24%和32%。预训练的俄语模型已经使用相同的神经网络架构对100小时的数据进行了训练。

英文摘要:

Development of Automatic Speech Recognition system for Kazakh language is very challenging due to a lack of data.Existing data of kazakh speech with its corresponding transcriptions are heavily accessed and not enough to gain a worth mentioning results.For this reason, speech recognition of Kazakh language has not been explored well.There are only few works that investigate this area with traditional methods Hidden Markov Model, Gaussian Mixture Model, but they are suffering from poor outcome and lack of enough data.In our work we suggest a new method that takes pre-trained model of Russian language and applies its knowledge as a starting point to our neural network structure, which means that we are transferring the weights of pre-trained model to our neural network.The main reason we chose Russian model is that pronunciation of kazakh and russian languages are quite similar because they share 78 percent letters and there are quite large corpus of russian speech dataset. We have collected a dataset of Kazakh speech with transcriptions in the base of Suleyman Demirel University with 50 native speakers each having around 400 sentences.Data have been chosen from famous Kazakh books. We have considered 4 different scenarios in our experiment. First, we trained our neural network without using a pre-trained Russian model with 2 LSTM layers and 2 BiLSTM .Second, we have trained the same 2 LSTM layered and 2 BiLSTM layered using a pre-trained model. As a result, we have improved our models training cost and Label Error Rate by using external Russian speech recognition model up to 24 percent and 32 percent respectively.Pre-trained Russian language model has trained on 100 hours of data with the same neural network architecture.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑